JP2005059656A - Landing control device and method for flying object - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device performing a landing control of a flying object having a new constitution regarding vertical speed control in landing to the place where the landing surface oscillates with time like a deck board of a ship. <P>SOLUTION: A detection part 10 detects a relative altitude from at least landing surface to the flying object. A parameter generation part 20 is a neural network outputting a landing control parameter by making a detection value detected by the detection part as an input, and is composed of the neural network having a feedback group inputted to a second node where an output of a first node among a plurality of nodes composing of the neural network is different from a first node. A control part 30 controls the flying object based on the control parameter outputted from the parameter generation part. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to an aircraft landing control apparatus and method, and more particularly to vertical speed control in landing on a place where a landing surface fluctuates with time, such as a ship deck.

  Conventionally, various control technologies using neural networks have been proposed as one of the technologies for controlling the controlled object with high accuracy regardless of the non-linear element, disturbance, or environmental change over time. Has been. In the neural network, necessary control parameters can be accurately output based on predetermined input information by appropriately setting the coupling weighting coefficient in the learning performed in advance. In learning in a neural network, a technique called a back propagation method is generally used. Examples of control using a neural network include air-fuel ratio control of an internal combustion engine disclosed in Patent Document 1, or control of a robot, machine tool, XY stage, disk device, etc. disclosed in Patent Document 2. .

By the way, a vertical take-off and landing aircraft such as a helicopter does not require a long runway during take-off and landing, and is used in many scenes because it does not choose a landing place, and automation of its control is desired.
JP-A-9-88685 JP-A-7-36506

  However, with this type of aircraft, the advantage of not choosing a landing location has made it difficult to automate control. For example, in the case of landing on the deck of a ship, the landing surface fluctuates due to the wavefront, so vertical speed control is more difficult than landing on a static landing surface. This is because if the speed control is mistaken, the impact at the time of landing will increase (hard landing state) and the aircraft may be damaged. In addition, since the wavefront has no regularity, it is difficult to perform control by predicting this, and it has been difficult to rely on the manual operation of a skilled operator.

  This invention is made | formed in view of this situation, The objective is to provide the landing control apparatus of the flying body which has a novel structure.

  Another object of the present invention is to appropriately perform vertical speed control when landing on a swinging landing surface.

  In order to solve such a problem, the first invention has at least a detection unit that detects a relative altitude from the landing surface to the flying object, and a detection value detected by the detection unit as inputs, and sets a landing control parameter of the flying object. A parameter generation unit configured with a neural network having a feedback loop in which an output of a first node in the neural network is input to a second node different from the first node; A landing control device for a flying object having a control unit for controlling the flying object based on a landing control parameter output from the unit.

  Here, in the first invention, the neural network preferably has a connection weight coefficient in the neural network learned by using a genetic algorithm. In this case, the coupling weight coefficient is obtained by evaluating the landing state of the flying object using a plurality of rocking models that reproduce the rocking of different landing surfaces. It is preferable that it is set. In addition, the plurality of swing models may be changed to different swing models for each evolution of a predetermined generation cycle. In addition, the coupling weight coefficient is determined as the optimal coupling weight coefficient that meets the evaluation conditions by evaluating the landing state of the flying object using an artificial rocking model in which the rocking state of the landing surface is intentionally manipulated by the operator. It is preferable that a solution is set.

  In the first invention, the neural network includes a plurality of layers including an input layer, an intermediate layer, and an output layer, and the output of the first node in the output layer is input to the input of the second node of the input layer. It is preferable to have a feedback loop for feedback.

  According to a second aspect of the present invention, there is provided a neural network having a feedback loop in which an output of a first node in a neural network is input to a second node different from the first node. A vehicle having a first step of outputting a landing control parameter of the flying object by inputting an altitude and a second step of controlling the flying object based on the landing control parameter output from the neural network. A landing control method is provided.

  Here, in the first invention or the second invention, it is preferable that the feedback loop delays the output of the first node and inputs it to the second node.

  In the second invention, it is preferable that the method further includes a third step of learning a connection weight coefficient in the neural network using a genetic algorithm. In this case, the third step evaluates the landing state of the flying object using a plurality of rocking models that reproduce the rocking of the landing surfaces that are different from each other. It is preferable that it is a step to set as. Further, the third step may include a step of changing a plurality of swing models to different swing models for each evolution of a predetermined generation period. Further, the third step is to evaluate the landing state of the flying object by further using an artificial rocking model in which the rocking state of the landing surface is intentionally operated by the operator, thereby obtaining an optimal solution that meets the evaluation condition. Preferably, the method includes a step of setting as a coupling weight coefficient.

  Thus, according to the present invention, a new landing control device can be provided by using a neural network. Also, by using a neural network having a feedback loop, the output of one node is fed back to another node. Thereby, the neural network can memorize the characteristics of the time-series transition related to the swing of the landing surface. Therefore, it is possible to stably control the landing by considering the time-series fluctuation of the landing surface in the past and predicting the state of the landing surface.

  FIG. 1 is a block configuration diagram of a landing control device for an aircraft according to the present embodiment. The landing control device 1 is mounted on a vertical take-off and landing aircraft (hereinafter simply referred to as “aircraft”) such as a helicopter and performs vertical speed control at the time of landing. The landing control device 1 includes a detection unit 10, a parameter generation unit 20, and a control unit 30.

  The detection unit 10 includes a first detection unit 11 that detects the state of the flying object, and a second detection unit 12 that detects a relative relationship between the flying object and the landing surface. Examples of the first detection unit 11 include an accelerometer, and the first detection unit 11 detects acceleration in the horizontal direction of the flying object (hereinafter referred to as “airframe acceleration”). Further, the first detection unit calculates the horizontal velocity of the flying object (hereinafter referred to as “aircraft speed”) by integrating the detected airframe acceleration. On the other hand, examples of the second detection unit include a stereo image processing device including a stereo camera and an image processing system. The second detection unit 12 allows the distance between the flying object and the landing surface (that is, relative altitude). ) Is detected. This stereo camera is attached to the lower part of the aircraft, and outputs a pair of captured images by imaging the scenery just below the aircraft. A pair of captured images output from the stereo camera is subjected to well-known stereo image processing by an image processing system provided at a later stage, thereby generating distance data. This distance data is a set of deviation amounts (that is, parallax) of minute image areas having a correlation in luminance between one captured image and the other captured image, and represents a two-dimensional distribution of distances below the aircraft. Show. Based on the distance data, the distance to the landing surface (relative altitude) is calculated. Further, based on the calculated amount of change in relative altitude per unit time, the velocity of the flying object in the vertical direction (hereinafter referred to as “vertical velocity”) is calculated. These values detected by the detection unit 10 are output to the subsequent parameter generation unit 20.

  FIG. 2 is an explanatory diagram showing details of the parameter generation unit 20. The parameter generation unit 20 is configured by a neural network NN. The neural network NN is a hierarchical neural network that includes a plurality of nodes N each having the same function, and each node is provided in a hierarchical manner (in this embodiment, it is composed of an input layer, an intermediate layer, and an output layer). Multi-layer hierarchical neural network). In this neural network NN, the number of nodes N constituting each layer is appropriately set by an operator. The operator considered the conflicting relationship (trade-off) between improving the reliability of the solution (output value) obtained by increasing the number of nodes N and improving the processing speed obtained by reducing the number of nodes N. Above, the network structure is determined so that each layer has the optimum number of nodes.

  In this neural network NN, nodes N corresponding to input information necessary for the neural network NN to generate and output landing control parameters (vertical landing speed in this embodiment) are set in the input layer. The In the present embodiment, the node N for inputting the airframe acceleration, the airframe speed, the relative altitude, and the vertical speed detected by the detection unit 10 is set in the input layer. In the output layer, a node N that outputs the landing control parameters of the flying object is set. On the other hand, the nodes N in the intermediate layer are prepared for the necessary number of nodes N based on the experience of the operator in consideration of the trade-off relationship described above.

Further, this neural network NN has a feedback loop in which the output of a certain node N (in this embodiment, the node N in the output layer) is input to another node N (in this embodiment, the node N in the input layer). Is provided. Therefore, the node N (one in this embodiment) for the purpose of output for feedback is further set in the output layer, and the node N (in this embodiment) for the purpose of input to this output is set in the input layer. In the form, three) are further set. By this loop, the output of the node N in the output layer is delayed by one cycle in system processing and then input to the node N in the input layer. Also, after the output of this node N is further delayed by one cycle by the first delay element Z ⁻¹ (thus after being delayed by two cycles from the initial input corresponding to the output of node N), Input to node N in the input layer. Further, the output from the first delay element Z ^-1, after being further delayed by one cycle by the second delay elements Z ^-1 provided at the subsequent stage (thus, the initial corresponding to the output node N After being delayed by three cycles from the input), it is input to node N in the input layer. In the present embodiment, this feedback loop refers to a loop existing in the neural network. That is, feedback from the output layer is completed within the neural network NN.

  Each node N performs calculations shown in Equations 1 and 2 for the input data yi, and outputs the calculation result as output data Yj. Here, wij is a coupling weight coefficient between the i-th element and the j-th element, and θj is a threshold value.

  As can be seen from these formulas 1 and 2, the output from a certain node N is input to another node N in accordance with a predetermined coupling weight coefficient Kij. Here, Formula 2 is called a sigmoid function and is generally used as a function of a node in the neural network NN. The sigmoid function changes continuously from 0 to 1, and approaches the step function as the threshold value θj decreases.

  When the landing control parameter is output using the neural network NN, it is necessary to appropriately adjust (learn) the coupling weight coefficient Kij and the threshold value θj in advance in order to improve the accuracy of the output result. Learning of these connection weight coefficients Kij and threshold value θj is performed using a genetic algorithm.

  The control unit 30 compares the landing control parameter (landing speed in the present embodiment) output from the parameter generation unit 20 with the current value of the landing control parameter (that is, the vertical speed). Then, various actuators are controlled so that the flying body has the landing speed generated and output by the parameter generation unit 20. Thereby, the engine output or the rotation speed / pitch of the rotor is controlled, and the vertical speed of the flying object is adjusted.

  FIG. 3 is a flowchart showing a procedure for determining the connection weight coefficient wij and the threshold value θj using a genetic algorithm. In this process, the flying object is controlled according to the landing control parameters output from the neural network NN while appropriately adjusting the values of the coupling weight coefficient wij and the threshold value θj by learning. Then, when it is determined that the behavior of the flying object has a predetermined evaluation condition, the learning ends. The evaluation conditions include “landing”, “low vertical velocity at landing”, “short time for landing”, “small amount of energy required for landing”, and the like. Note that if the flying object is directly controlled, there is a risk of destroying the actual aircraft, so learning proceeds by simulation (see FIG. 4).

  First, in step 1, an initial individual group in which a plurality of individuals composed of genotypes are collected is generated. The genotype corresponds to the coupling weight coefficient Kij and the threshold value θj. One individual has all the connection weight coefficients Kij and threshold values θj in the neural network NN as genotypes. The initial individual population is a collection of n such individuals, and is composed of genotypes Kij and θj of various values for each individual. The initial values of the genotypes Kij and θj constituting each individual 1 to n are determined by random numbers, for example. Therefore, if the connection weight coefficient Kij and the threshold value θj are set in the neural network NN for each of the individuals 1 to n, n neural networks NN in which the connection weight coefficient Kij and the threshold value θj are set are obtained.

  In step 2, a landing simulation is performed. In this simulation, the landing state at the time of controlling the flying object is reproduced based on the landing control parameters output from the neural network NN in which the genotypes Kij and θj are set for each individual 1 to N. The The simulator 40 is provided with a swing model that expresses a swing state of a landing surface on a ship according to a wave front state by a well-known function formula. Since there are various variations in the swinging state of the landing surface, a plurality of (for example, 20) swing models each reproducing the swinging of the different landing surfaces in one generation of learning for each individual 1 to n. Is used. The simulator 40 is also provided with an artificial rocking model (specifically, a dynamic rocking state of the landing surface that cannot be reproduced by rocking the landing surface due to the wavefront) that is intentionally operated by the operator. Has been. FIG. 5 is a diagram illustrating an example of an artificial swing model. In the figure, the time course of the landing surface (oscillation pattern) reproduced by the artificial rocking model is shown by a solid line, and the time course of the relative altitude of the flying object is shown by a dotted line. . In this artificial rocking model, the landing surface is kept stationary until the distance between the flying object and the landing surface becomes a certain value (for example, 1 m) or less. When the flying object comes closer to the landing surface than this fixed value, the landing surface moves downward (in a direction away from the flying object) by a predetermined amount. Then, after moving downward by a certain amount, the landing surface moves upward (in the direction approaching the flying object) until the flying object lands.

  The landing state is reproduced based on the behavior of the flying object reproduced by the flying object model with respect to the swing model and the artificial swing model. This flying object model is obtained by modeling the flying object behavior in advance using a functional equation, and thereby, the actual behavior of the flying object when the flying object is controlled based on a certain landing control parameter is reproduced. In this simulation, the relative height and vertical velocity of the flying object are calculated from the simulator 40 and fed back to the neural network NN. Further, the aircraft velocity and acceleration of the aircraft are calculated from the aircraft model, and these values are also fed back to the neural network NN. By repeating such processing, a landing state simulation for the swing model (or artificial swing model) is performed. FIG. 6 is a diagram showing the altitude trajectory of the flying object in the landing simulation. In the figure, the time course of the relative altitude of the flying object is shown by a solid line, and the swing pattern of the landing surface is drawn by a dotted line. Through the simulation, it is determined whether the flying object has landed, the speed of the aircraft at the time of landing, the time required for landing, and the amount of energy required for landing.

  In Step 3, an evaluation value A is calculated for each individual 1 to n based on the simulation result. This evaluation value A has a plurality of evaluation conditions (compounds such as “landing”, “low aircraft speed at landing”, “short time required for landing”, and “low amount of energy required for landing”). The value is smaller as the individual satisfies the condition), that is, as the individual obtains a good evaluation. For example, the evaluation value A is calculated individually for each evaluation condition, and the product-sum operation is performed by setting a weighting factor according to the importance of the evaluation condition corresponding to the calculated evaluation value. And so on. Then, each individual 1 to n in the individual group is newly rearranged as the individual 1 to n in order from the individual having the smallest evaluation value A.

  In step 4, it is determined whether or not the evaluation value A related to the individual 1 is equal to or less than the determination evaluation value Aerror. This judgment evaluation value Aerror is set in advance through experiments and simulations as the maximum value of the evaluation value A that can be considered that the landing state of the flying vehicle sufficiently satisfies the above-mentioned composite condition. If a negative determination is made in step 4, that is, if the evaluation value A is larger than the determination evaluation value A error, the process proceeds to the subsequent step 5. On the other hand, if an affirmative determination is made in step 4, that is, if the evaluation value A is equal to or less than the determination evaluation value A error, the process proceeds to step 8.

  In step 5, an evolutionary operation is performed using a genetic algorithm. Specifically, selection and selection of individuals 1 to n are performed in the individual population. Examples of such selection and selection methods include reverse roulette type selection, rank type selection, tournament type selection, and the like. A certain individual (or individual group) is selected by the processing of step 4, and the selected individual is deleted from the individual group. For example, by moving the same number of individuals from the smaller evaluation value A to the positions of the deleted individuals, the number of individuals constituting the individual population is maintained. Next, a next generation individual population is generated. Specifically, the optimization unit 10 mutates and crosses the genotypes Kij and θj constituting the individuals in the group. In the mutation, an arbitrary genotype Kij (or θj) in an individual is selected by, for example, a random number and changed to a value generated by the random number. In crossover, in a certain group selected by random numbers, the values of genotypes Kij (or θj) also selected by random numbers are exchanged with each other. However, the method of selecting individuals to be mutated and crossed is not limited to selecting by random numbers. For the individual 1 having the smallest evaluation value A, mutation and crossover are not performed in order to maintain the individual. You may do it.

  In step 6, it is determined whether or not the current individual population has evolved for a predetermined generation period (for example, 100 generations). The reason for providing this determination is that overlearning occurs when learning is performed only under certain specific conditions. Overlearning gives excellent results under the learned conditions (20 fluctuation models), but it cannot obtain similar results for unknown conditions due to the influence of disturbances. That is. For this reason, it is necessary to change the condition (oscillation model) at an appropriate period to suppress the occurrence of overlearning. If an affirmative determination is made in step 6, that is, if a certain generation period has evolved (for example, n × 100 generations (n = 1, 2,...)), The process proceeds to step 7. On the other hand, if a negative determination is made in this step, step 7 is skipped and the process returns to step 2.

  In step 7, the swing model is changed. For this change, all 20 swing models may be changed, or some swing models (for example, about 10) may be selectively changed. Then, returning to the processing after step 2, the above-described processing is repeated until the evaluation value A of the individual 1 becomes equal to or less than the determination evaluation value Aerror.

  In step 8, the genotypes Kij and θj constituting the individual 1, that is, the genotypes Kij and θj that match the evaluation conditions are determined as the optimum solutions, and the routine is exited. In this case, based on the genotypes Kij and θj constituting the individual 1, the values of these genotypes Kij and θj are set as the connection weight coefficient Kij and the threshold value θj in the neural network NN.

  FIG. 7 is a diagram showing the transition of the evaluation value A with the generation evolution. As can be seen from the figure, the evaluation value A decreases as the generation progresses. This means that learning progresses as the generation progresses, and the connection weight coefficient Kij and threshold value θj of the neural network NN are appropriately adjusted. When a certain generation period is reached, the swing model is changed. Therefore, an unknown condition is newly added, and the evaluation value A temporarily increases. However, the evaluation value A gradually decreases as the generation progresses and learning progresses. If the swing model is changed periodically, the evaluation value temporarily deteriorates. However, by repeating the change, the response to various swing models increases, and the evaluation value when the swing model is changed. The rising width of A (the degree of increase) gradually decreases. Finally, learning is performed until the evaluation value A becomes equal to or less than the determination evaluation value Aerror, and optimal genotypes Kij and θj can be obtained.

  As described above, according to the present embodiment, the genetic weight is used to learn the connection weight coefficient Kij (and the threshold value θj) of the neural network NN having the feedback loop, and the optimum solution is determined. ing. The neural network NN having feedback has a problem that a learning rule based on the principle of steepest descent method such as backpropagation cannot be performed. However, in the present embodiment, this can be solved by using a genetic algorithm.

  Further, by using the neural network NN having feedback, the output of the node N in the output layer is fed back to the node N in the input layer. In the neural network NN having no feedback, the output of the output layer (landing control parameter) depends on the input at a certain point in time, and the past input history does not affect the output. Therefore, even if time-series data is input, no output is performed in consideration of the characteristics of the transition of the previous time-series data. In this embodiment, the characteristics of the time-series transition (swing pattern) related to the swing of the landing surface are stored by the structure of the feedback loop of the neural network. As a result, past time-series fluctuations of the landing surface are taken into consideration, and landing control parameters foreseeing the state of the landing surface are output by the neural network NN. For this reason, stable speed control can be performed by using this landing control parameter.

  Note that the above-described structure of the neural network NN is merely an example, and besides this, a node N for the purpose of feedback output may be further provided in the output layer. Further, the output layer may be configured with only a single node for the purpose of landing control parameter output and feedback output. Furthermore, if the feedback is accompanied by a time delay, the feedback loop may be provided from the output layer to the input layer, or from the intermediate layer to the input layer, or from the output layer to the intermediate layer. There are various variations.

  Further, in this embodiment, when learning the coupling weight coefficient Kij and the threshold value θj, the swing model is changed at a predetermined generation cycle. According to this method, even if the number of swing models to be evaluated is initially small, evaluation can be performed using a large number of swing models by periodically changing the swing pattern. Can do. Thereby, overlearning can be suppressed by the diversity of the swing model. In view of suppressing overlearning, it is conceivable to perform evaluation using a large number of swing patterns. However, with this method, the solution space is widened and it is difficult to obtain an optimal solution. Therefore, there is a problem that the convergence to the judgment evaluation value Aerror is prolonged. In this regard, in this embodiment, the evaluation time is converged to the judgment evaluation value Aerror by performing evaluation while evaluating with a relatively small number of swing patterns and selecting excellent individuals. Can be suppressed.

  Further, when an individual is evaluated, by using an artificial rocking model set under conditions more severe than the rocking model, it is possible to learn in advance the ability to cope with sudden changes that may occur in an actual environment. In addition, since the swing model is changed periodically, there is an advantage that the result can be commonly used as an evaluation guideline under a certain condition by evaluating the artificial swing model for any generation of individuals.

  When learning by a genetic algorithm, it is often difficult to escape from a local solution by providing a plurality of evaluation conditions. For this reason, a complicated compound condition may cause a phenomenon that the calculation time required for the solution search becomes extremely slow or does not advance at all (so-called trap to a local solution). Therefore, in order to solve such a problem, an optimization method using a so-called stepwise evolution method may be used. In this stepwise evolution method, a genetic algorithm is used to calculate an optimal solution that matches an evaluation condition in which a plurality of conditions are combined. Specifically, based on a genetic algorithm, a GA engine that outputs solution candidates that meet individually set evaluation conditions as an evolutionary population PGn (n = 1 to N) is provided from the lower level to the higher level. Multiple stages are provided. The individual immigration control unit moves a part of the evolved populations PGn and PGn-1 between the lower-order evolved population PGn and the higher-order evolved population PGn + 1. The convergence determination unit performs optimization convergence determination based on the evolution population PGN of the highest GA engine. The evaluation conditions set in each stage are added step by step from the lower level to the higher level. Thereby, the evolution calculation according to the optimization degree of the individual is automatically and autonomously performed. In addition, the solution search for the evolutionary population is performed efficiently, and the bias of individual genes can be avoided. As a result, the possibility of being trapped by the local solution is reduced, and the progress of evolution becomes more continuous, so that a high-quality solution can be calculated at high speed. The details of the optimization method using the stepwise evolution method are disclosed in Japanese Patent Laid-Open No. 2002-31755, so refer to them if necessary.

  Further, the state detected by the detection unit 10 is not limited to the above form. For example, the first detector 11 may detect the attitude angular velocity of the flying object using a gyro and may integrate the attitude angular velocity to calculate the attitude angle of the flying object. As a result, the information input to the parameter generation unit 20 can be enriched, so that more accurate control can be performed. Further, the second detection unit may calculate a relative altitude using a distance measuring sensor such as a laser radar or a millimeter wave radar. If the relationship between the flying object and the landing surface is detected, information may be acquired from the landing surface side to detect the relative relationship between the two. However, in this case, it is necessary to share information between the landing surface and the flying object, and it takes a cost to install the equipment, and there is a risk that the landing surface is limited.

  Further, the output from the neural network NN is not limited to the landing speed, and may be an optimum relative altitude or the like. Moreover, you may output directly the control value which controls a flying body, such as an engine output value and the rotation speed / pitch of a rotor. In addition, the landing control device according to the present embodiment is not limited to landing on the deck surface on the ship, for example, landing on a vehicle traveling on a road with undulations, landing on a floating body at sea, etc. There are various. Needless to say, the present invention can also be applied to control in the vertical direction with respect to a static landing surface.

Block configuration diagram of the landing control device according to the present embodiment Explanatory drawing showing details of parameter generator A flowchart showing the procedure for determining the connection weighting factor and threshold using a genetic algorithm Diagram showing system configuration including simulator Diagram showing an example of an artificial swing model Figure showing the altitude trajectory of the flying object in the landing simulation Figure showing the transition of evaluation values as generations evolve

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Landing control apparatus 10 Detection part 11 1st detection part 12 2nd detection part 20 Parameter generation part 30 Control part 40 Simulator NN Neural network

Claims (13)

In the landing control device of the flying object,
A detection unit that detects at least the relative altitude from the landing surface to the flying object;
A neural network that outputs a landing control parameter of the flying object by using a detection value detected by the detection unit as an input, wherein the output of the first node in the neural network is different from the first node. A parameter generator composed of the neural network having a feedback loop input to the nodes of
A flying object landing control apparatus comprising: a control unit that controls the flying object based on the landing control parameter output from the parameter generation unit.   The aircraft landing control apparatus according to claim 1, wherein the feedback loop delays the output of the first node and inputs the output to the second node.   3. The flying object landing control apparatus according to claim 1, wherein the neural network has learned a connection weight coefficient in the neural network using a genetic algorithm.   The coupling weight coefficient is obtained by evaluating the landing state of the flying object using a plurality of rocking models that reproduce the rocking of the landing surface, which are different from each other. The landing control device for a flying object according to claim 3, wherein:   5. The flying object landing control apparatus according to claim 4, wherein the plurality of rocking models are changed from the rocking models different for each evolution of a predetermined generation period.   The coupling weight coefficient is obtained by evaluating the landing state of the flying object using an artificial rocking model in which the rocking state of the landing surface is intentionally operated by an operator, so that the coupling weight that satisfies the evaluation condition is used. 6. The flying object landing control apparatus according to claim 4 or 5, wherein an optimum solution of the coefficient is set.   The neural network includes a plurality of layers including an input layer, an intermediate layer, and an output layer, and feeds back an output of the first node in the output layer to an input of the second node of the input layer. 7. The flying object landing control apparatus according to claim 1, further comprising a feedback loop. In the landing control method of the aircraft,
By inputting at least the relative altitude from the landing surface to the flying object into the neural network having a feedback loop in which the output of the first node in the neural network is input to a second node different from the first node. A first step of outputting landing control parameters of the aircraft;
And a second step of controlling the flying object based on the landing control parameter output from the neural network.   9. The aircraft landing control method according to claim 8, wherein the feedback loop delays the output of the first node and inputs it to the second node.   10. The flying object landing control method according to claim 8, further comprising a third step of learning a connection weight coefficient in the neural network using a genetic algorithm.   The third step evaluates the landing state of the flying object using a plurality of swing models that reproduce the swings of the landing surfaces that are different from each other, thereby obtaining an optimal solution that meets the evaluation conditions. 11. The flying object landing control method according to claim 10, wherein the step is set as a coefficient.   12. The vehicle landing control according to claim 11, wherein the third step includes a step of changing a plurality of rocking models with different rocking models for each evolution of a predetermined generation period. Method.   In the third step, the landing state of the flying object is further evaluated by using an artificial rocking model in which the rocking state of the landing surface is intentionally operated by an operator. The method according to claim 11 or 12, further comprising the step of setting a solution as the coupling weight coefficient.

Images